Meta’s MM-slotgate boosts fashion AI search by 56%
MM-slotgate splits fashion search into 4 controllable attributes, lifting color accuracy from 32% to 89%
Meta’s AI team has unveiled MM-slotgate, a novel approach to fashion retrieval that moves beyond monolithic embeddings by factorizing queries into four named attribute slots: color, pattern, category, and demographic. Each slot is paired with a learned gating mechanism that dynamically weights image vs. text evidence based on attribute type—color retrieval leans 57.4% on image data, while category relies more on text.
In tests on H&M’s dataset, MM-slotgate achieved a 0.7566 macro ConstraintSatisfied@10 score, outperforming equal-weight multimodal fusion (0.7142) and text-only fCLIP (0.4755). The most dramatic improvement came in color search, where accuracy jumped from 32.1% to 88.9% (+56.8%). The model’s interpretable gates require no external supervision, and its quantized slot codes enable targeted interventions, such as a 15.3x lift for color correction. The system demonstrates that typed, attribute-conditioned multimodal slots outperform both global embeddings and opaque semantic IDs for controllable fashion retrieval.
- MM-slotgate splits fashion search into 4 interpretable attribute slots (color, pattern, category, demographic) with dynamic image/text weighting
- Achieves 0.7566 ConstraintSatisfied@10 on H&M data—6% better than prior methods—and improves color accuracy by 56% (0.321→0.889)
- Learned gates are unsupervised and interpretable; quantized slots enable targeted interventions like a 15.3x lift for color correction
Why It Matters
Enables e-commerce platforms to offer precise, attribute-aware fashion search at scale, cutting irrelevant results and boosting conversion.